Skip to content

[FEA] Beam search optimization #277

@JacoCheung

Description

@JacoCheung

We need to

  1. Support target group attention (This can be addressed in [FEA] Jagged Arbitrary Masked Self Attention support #276)
  2. Support dynamic beam width.
  3. KVCache usage.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementImprovement for existing featurefeatureNew feature

    Type

    No type

    Projects

    No projects

    Milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions